Improved handling of asynchronous copy, scheduling and synchronization for A100. Now achieving CUTLASS-like performance on large square dense matrix multiplication tasks
- A100 support via mma.16816
- Thread swizzling for conflict-free shared memory accesses without
padding
- Complete overhaul of the LLVM code generation in
codegen/selection/generator.cc to remove overengineering
- Added debugging capabilities in the Python binding
- Compilation error for kernels that spill
improvements and bugfixes:
* Added preliminary support for extended Einstein summation in PyTriton
* Significant performance improvement on FP32 kernels containing matrix
multiplication
* Added re-coalescing pass for FP16 kernels containing matrix
multiplication
* Various bugfixes
